The Effect of Learning Rate and Regularization on the Performance of ResNet50 for Classifying Palembang Woven Fabric Motifs

Abstract

Palembang woven fabric is one of Indonesia's cultural heritages, characterized by diverse motifs with distinctive visual patterns that require accurate identification methods to support cultural preservation and digitalization. This study aims to analyze the effect of learning rate and regularization techniques on the performance of a transfer learning-based ResNet50 model for Palembang woven fabric motif classification. The dataset consisted of 800 images representing eight motif classes, divided into 70% training, 15% validation, and 15% testing sets. All images underwent preprocessing, including resizing to 224 × 224 pixels and normalization. The ResNet50 model was trained using the Adam optimizer for 50 epochs with three learning rates (0.01, 0.001, and 0.0001) and two regularization techniques, namely L2 Regularization and Dropout. Model performance was evaluated using Accuracy, Precision, Recall, and F1-Score, while training and validation curves were analyzed to assess model convergence. The experimental results demonstrate that both learning rate and regularization techniques significantly influence classification performance. The best performance was achieved using a learning rate of 0.001 with Dropout, resulting in a training accuracy of 99.11%, validation accuracy of 96.67%, and testing accuracy of 95.83%, outperforming all other configurations.